Merge branch 'TRUNK' into ARE-5.0

1 files changed, 181 insertions, 181 deletions

diff --git a/libs/ffvpx/libavcodec/vp9dsp_template.c b/libs/ffvpx/libavcodec/vp9dsp_template.c
index bb54561a6..9b1166170 100644
--- a/libs/ffvpx/libavcodec/vp9dsp_template.c
+++ b/libs/ffvpx/libavcodec/vp9dsp_template.c
@@ -1138,7 +1138,7 @@ static void type_a##_##type_b##_##sz##x##sz##_add_c(uint8_t *_dst, \
             for (j = 0; j < sz; j++) \
                 dst[j * stride] = av_clip_pixel(dst[j * stride] + \
                                                 (bits ? \
-                                                 (t + (1 << (bits - 1))) >> bits : \
+                                                 (int)(t + (1U << (bits - 1))) >> bits : \
                                                  t)); \
             dst++; \
         } \
@@ -1153,7 +1153,7 @@ static void type_a##_##type_b##_##sz##x##sz##_add_c(uint8_t *_dst, \
         for (j = 0; j < sz; j++) \
             dst[j * stride] = av_clip_pixel(dst[j * stride] + \
                                             (bits ? \
-                                             (out[j] + (1 << (bits - 1))) >> bits : \
+                                             (int)(out[j] + (1U << (bits - 1))) >> bits : \
                                              out[j])); \
         dst++; \
     } \
@@ -1260,25 +1260,25 @@ static av_always_inline void iadst8_1d(const dctcoef *in, ptrdiff_t stride,
     t6 = (t2a - t6a + (1 << 13)) >> 14;
     t7 = (t3a - t7a + (1 << 13)) >> 14;
 
-    t4a = 15137 * t4 +  6270 * t5;
-    t5a =  6270 * t4 - 15137 * t5;
-    t6a = 15137 * t7 -  6270 * t6;
-    t7a =  6270 * t7 + 15137 * t6;
+    t4a = 15137U * t4 +  6270U * t5;
+    t5a =  6270U * t4 - 15137U * t5;
+    t6a = 15137U * t7 -  6270U * t6;
+    t7a =  6270U * t7 + 15137U * t6;
 
     out[0] =   t0 + t2;
     out[7] = -(t1 + t3);
     t2     =   t0 - t2;
     t3     =   t1 - t3;
 
-    out[1] = -((t4a + t6a + (1 << 13)) >> 14);
-    out[6] =   (t5a + t7a + (1 << 13)) >> 14;
-    t6     =   (t4a - t6a + (1 << 13)) >> 14;
-    t7     =   (t5a - t7a + (1 << 13)) >> 14;
+    out[1] = -((dctint)((1U << 13) + t4a + t6a) >> 14);
+    out[6] =   (dctint)((1U << 13) + t5a + t7a) >> 14;
+    t6     =   (dctint)((1U << 13) + t4a - t6a) >> 14;
+    t7     =   (dctint)((1U << 13) + t5a - t7a) >> 14;
 
-    out[3] = -(((t2 + t3) * 11585 + (1 << 13)) >> 14);
-    out[4] =   ((t2 - t3) * 11585 + (1 << 13)) >> 14;
-    out[2] =   ((t6 + t7) * 11585 + (1 << 13)) >> 14;
-    out[5] = -(((t6 - t7) * 11585 + (1 << 13)) >> 14);
+    out[3] = -((dctint)((t2 + t3) * 11585U + (1 << 13)) >> 14);
+    out[4] =   (dctint)((t2 - t3) * 11585U + (1 << 13)) >> 14;
+    out[2] =   (dctint)((t6 + t7) * 11585U + (1 << 13)) >> 14;
+    out[5] = -((dctint)((t6 - t7) * 11585U + (1 << 13)) >> 14);
 }
 
 itxfm_wrap(8, 5)
@@ -1290,22 +1290,22 @@ static av_always_inline void idct16_1d(const dctcoef *in, ptrdiff_t stride,
     dctint t0a, t1a, t2a, t3a, t4a, t5a, t6a, t7a;
     dctint t8a, t9a, t10a, t11a, t12a, t13a, t14a, t15a;
 
-    t0a  = ((IN(0) + IN(8)) * 11585 + (1 << 13)) >> 14;
-    t1a  = ((IN(0) - IN(8)) * 11585 + (1 << 13)) >> 14;
-    t2a  = (IN(4)  *  6270 - IN(12) * 15137 + (1 << 13)) >> 14;
-    t3a  = (IN(4)  * 15137 + IN(12) *  6270 + (1 << 13)) >> 14;
-    t4a  = (IN(2)  *  3196 - IN(14) * 16069 + (1 << 13)) >> 14;
-    t7a  = (IN(2)  * 16069 + IN(14) *  3196 + (1 << 13)) >> 14;
-    t5a  = (IN(10) * 13623 - IN(6)  *  9102 + (1 << 13)) >> 14;
-    t6a  = (IN(10) *  9102 + IN(6)  * 13623 + (1 << 13)) >> 14;
-    t8a  = (IN(1)  *  1606 - IN(15) * 16305 + (1 << 13)) >> 14;
-    t15a = (IN(1)  * 16305 + IN(15) *  1606 + (1 << 13)) >> 14;
-    t9a  = (IN(9)  * 12665 - IN(7)  * 10394 + (1 << 13)) >> 14;
-    t14a = (IN(9)  * 10394 + IN(7)  * 12665 + (1 << 13)) >> 14;
-    t10a = (IN(5)  *  7723 - IN(11) * 14449 + (1 << 13)) >> 14;
-    t13a = (IN(5)  * 14449 + IN(11) *  7723 + (1 << 13)) >> 14;
-    t11a = (IN(13) * 15679 - IN(3)  *  4756 + (1 << 13)) >> 14;
-    t12a = (IN(13) *  4756 + IN(3)  * 15679 + (1 << 13)) >> 14;
+    t0a  = (dctint)((IN(0) + IN(8)) * 11585U + (1 << 13)) >> 14;
+    t1a  = (dctint)((IN(0) - IN(8)) * 11585U + (1 << 13)) >> 14;
+    t2a  = (dctint)(IN(4)  *  6270U - IN(12) * 15137U + (1 << 13)) >> 14;
+    t3a  = (dctint)(IN(4)  * 15137U + IN(12) *  6270U + (1 << 13)) >> 14;
+    t4a  = (dctint)(IN(2)  *  3196U - IN(14) * 16069U + (1 << 13)) >> 14;
+    t7a  = (dctint)(IN(2)  * 16069U + IN(14) *  3196U + (1 << 13)) >> 14;
+    t5a  = (dctint)(IN(10) * 13623U - IN(6)  *  9102U + (1 << 13)) >> 14;
+    t6a  = (dctint)(IN(10) *  9102U + IN(6)  * 13623U + (1 << 13)) >> 14;
+    t8a  = (dctint)(IN(1)  *  1606U - IN(15) * 16305U + (1 << 13)) >> 14;
+    t15a = (dctint)(IN(1)  * 16305U + IN(15) *  1606U + (1 << 13)) >> 14;
+    t9a  = (dctint)(IN(9)  * 12665U - IN(7)  * 10394U + (1 << 13)) >> 14;
+    t14a = (dctint)(IN(9)  * 10394U + IN(7)  * 12665U + (1 << 13)) >> 14;
+    t10a = (dctint)(IN(5)  *  7723U - IN(11) * 14449U + (1 << 13)) >> 14;
+    t13a = (dctint)(IN(5)  * 14449U + IN(11) *  7723U + (1 << 13)) >> 14;
+    t11a = (dctint)(IN(13) * 15679U - IN(3)  *  4756U + (1 << 13)) >> 14;
+    t12a = (dctint)(IN(13) *  4756U + IN(3)  * 15679U + (1 << 13)) >> 14;
 
     t0  = t0a  + t3a;
     t1  = t1a  + t2a;
@@ -1324,12 +1324,12 @@ static av_always_inline void idct16_1d(const dctcoef *in, ptrdiff_t stride,
     t14 = t15a - t14a;
     t15 = t15a + t14a;
 
-    t5a  = ((t6 - t5) * 11585 + (1 << 13)) >> 14;
-    t6a  = ((t6 + t5) * 11585 + (1 << 13)) >> 14;
-    t9a  = (  t14 *  6270 - t9  * 15137  + (1 << 13)) >> 14;
-    t14a = (  t14 * 15137 + t9  *  6270  + (1 << 13)) >> 14;
-    t10a = (-(t13 * 15137 + t10 *  6270) + (1 << 13)) >> 14;
-    t13a = (  t13 *  6270 - t10 * 15137  + (1 << 13)) >> 14;
+    t5a  = (dctint)((t6 - t5) * 11585U + (1 << 13)) >> 14;
+    t6a  = (dctint)((t6 + t5) * 11585U + (1 << 13)) >> 14;
+    t9a  = (dctint)(  t14 *  6270U - t9  * 15137U  + (1 << 13)) >> 14;
+    t14a = (dctint)(  t14 * 15137U + t9  *  6270U  + (1 << 13)) >> 14;
+    t10a = (dctint)(-(t13 * 15137U + t10 *  6270U) + (1 << 13)) >> 14;
+    t13a = (dctint)(  t13 *  6270U - t10 * 15137U  + (1 << 13)) >> 14;
 
     t0a  = t0   + t7;
     t1a  = t1   + t6a;
@@ -1348,10 +1348,10 @@ static av_always_inline void idct16_1d(const dctcoef *in, ptrdiff_t stride,
     t14  = t14a + t13a;
     t15a = t15  + t12;
 
-    t10a = ((t13  - t10)  * 11585 + (1 << 13)) >> 14;
-    t13a = ((t13  + t10)  * 11585 + (1 << 13)) >> 14;
-    t11  = ((t12a - t11a) * 11585 + (1 << 13)) >> 14;
-    t12  = ((t12a + t11a) * 11585 + (1 << 13)) >> 14;
+    t10a = (dctint)((t13  - t10)  * 11585U + (1 << 13)) >> 14;
+    t13a = (dctint)((t13  + t10)  * 11585U + (1 << 13)) >> 14;
+    t11  = (dctint)((t12a - t11a) * 11585U + (1 << 13)) >> 14;
+    t12  = (dctint)((t12a + t11a) * 11585U + (1 << 13)) >> 14;
 
     out[ 0] = t0a + t15a;
     out[ 1] = t1a + t14;
@@ -1378,48 +1378,48 @@ static av_always_inline void iadst16_1d(const dctcoef *in, ptrdiff_t stride,
     dctint t0a, t1a, t2a, t3a, t4a, t5a, t6a, t7a;
     dctint t8a, t9a, t10a, t11a, t12a, t13a, t14a, t15a;
 
-    t0  = IN(15) * 16364 + IN(0)  *   804;
-    t1  = IN(15) *   804 - IN(0)  * 16364;
-    t2  = IN(13) * 15893 + IN(2)  *  3981;
-    t3  = IN(13) *  3981 - IN(2)  * 15893;
-    t4  = IN(11) * 14811 + IN(4)  *  7005;
-    t5  = IN(11) *  7005 - IN(4)  * 14811;
-    t6  = IN(9)  * 13160 + IN(6)  *  9760;
-    t7  = IN(9)  *  9760 - IN(6)  * 13160;
-    t8  = IN(7)  * 11003 + IN(8)  * 12140;
-    t9  = IN(7)  * 12140 - IN(8)  * 11003;
-    t10 = IN(5)  *  8423 + IN(10) * 14053;
-    t11 = IN(5)  * 14053 - IN(10) *  8423;
-    t12 = IN(3)  *  5520 + IN(12) * 15426;
-    t13 = IN(3)  * 15426 - IN(12) *  5520;
-    t14 = IN(1)  *  2404 + IN(14) * 16207;
-    t15 = IN(1)  * 16207 - IN(14) *  2404;
-
-    t0a  = (t0 + t8  + (1 << 13)) >> 14;
-    t1a  = (t1 + t9  + (1 << 13)) >> 14;
-    t2a  = (t2 + t10 + (1 << 13)) >> 14;
-    t3a  = (t3 + t11 + (1 << 13)) >> 14;
-    t4a  = (t4 + t12 + (1 << 13)) >> 14;
-    t5a  = (t5 + t13 + (1 << 13)) >> 14;
-    t6a  = (t6 + t14 + (1 << 13)) >> 14;
-    t7a  = (t7 + t15 + (1 << 13)) >> 14;
-    t8a  = (t0 - t8  + (1 << 13)) >> 14;
-    t9a  = (t1 - t9  + (1 << 13)) >> 14;
-    t10a = (t2 - t10 + (1 << 13)) >> 14;
-    t11a = (t3 - t11 + (1 << 13)) >> 14;
-    t12a = (t4 - t12 + (1 << 13)) >> 14;
-    t13a = (t5 - t13 + (1 << 13)) >> 14;
-    t14a = (t6 - t14 + (1 << 13)) >> 14;
-    t15a = (t7 - t15 + (1 << 13)) >> 14;
-
-    t8   = t8a  * 16069 + t9a  *  3196;
-    t9   = t8a  *  3196 - t9a  * 16069;
-    t10  = t10a *  9102 + t11a * 13623;
-    t11  = t10a * 13623 - t11a *  9102;
-    t12  = t13a * 16069 - t12a *  3196;
-    t13  = t13a *  3196 + t12a * 16069;
-    t14  = t15a *  9102 - t14a * 13623;
-    t15  = t15a * 13623 + t14a *  9102;
+    t0  = IN(15) * 16364U + IN(0)  *   804U;
+    t1  = IN(15) *   804U - IN(0)  * 16364U;
+    t2  = IN(13) * 15893U + IN(2)  *  3981U;
+    t3  = IN(13) *  3981U - IN(2)  * 15893U;
+    t4  = IN(11) * 14811U + IN(4)  *  7005U;
+    t5  = IN(11) *  7005U - IN(4)  * 14811U;
+    t6  = IN(9)  * 13160U + IN(6)  *  9760U;
+    t7  = IN(9)  *  9760U - IN(6)  * 13160U;
+    t8  = IN(7)  * 11003U + IN(8)  * 12140U;
+    t9  = IN(7)  * 12140U - IN(8)  * 11003U;
+    t10 = IN(5)  *  8423U + IN(10) * 14053U;
+    t11 = IN(5)  * 14053U - IN(10) *  8423U;
+    t12 = IN(3)  *  5520U + IN(12) * 15426U;
+    t13 = IN(3)  * 15426U - IN(12) *  5520U;
+    t14 = IN(1)  *  2404U + IN(14) * 16207U;
+    t15 = IN(1)  * 16207U - IN(14) *  2404U;
+
+    t0a  = (dctint)((1U << 13) + t0 + t8 ) >> 14;
+    t1a  = (dctint)((1U << 13) + t1 + t9 ) >> 14;
+    t2a  = (dctint)((1U << 13) + t2 + t10) >> 14;
+    t3a  = (dctint)((1U << 13) + t3 + t11) >> 14;
+    t4a  = (dctint)((1U << 13) + t4 + t12) >> 14;
+    t5a  = (dctint)((1U << 13) + t5 + t13) >> 14;
+    t6a  = (dctint)((1U << 13) + t6 + t14) >> 14;
+    t7a  = (dctint)((1U << 13) + t7 + t15) >> 14;
+    t8a  = (dctint)((1U << 13) + t0 - t8 ) >> 14;
+    t9a  = (dctint)((1U << 13) + t1 - t9 ) >> 14;
+    t10a = (dctint)((1U << 13) + t2 - t10) >> 14;
+    t11a = (dctint)((1U << 13) + t3 - t11) >> 14;
+    t12a = (dctint)((1U << 13) + t4 - t12) >> 14;
+    t13a = (dctint)((1U << 13) + t5 - t13) >> 14;
+    t14a = (dctint)((1U << 13) + t6 - t14) >> 14;
+    t15a = (dctint)((1U << 13) + t7 - t15) >> 14;
+
+    t8   = t8a  * 16069U + t9a  *  3196U;
+    t9   = t8a  *  3196U - t9a  * 16069U;
+    t10  = t10a *  9102U + t11a * 13623U;
+    t11  = t10a * 13623U - t11a *  9102U;
+    t12  = t13a * 16069U - t12a *  3196U;
+    t13  = t13a *  3196U + t12a * 16069U;
+    t14  = t15a *  9102U - t14a * 13623U;
+    t15  = t15a * 13623U + t14a *  9102U;
 
     t0   = t0a + t4a;
     t1   = t1a + t5a;
@@ -1429,49 +1429,49 @@ static av_always_inline void iadst16_1d(const dctcoef *in, ptrdiff_t stride,
     t5   = t1a - t5a;
     t6   = t2a - t6a;
     t7   = t3a - t7a;
-    t8a  = (t8  + t12 + (1 << 13)) >> 14;
-    t9a  = (t9  + t13 + (1 << 13)) >> 14;
-    t10a = (t10 + t14 + (1 << 13)) >> 14;
-    t11a = (t11 + t15 + (1 << 13)) >> 14;
-    t12a = (t8  - t12 + (1 << 13)) >> 14;
-    t13a = (t9  - t13 + (1 << 13)) >> 14;
-    t14a = (t10 - t14 + (1 << 13)) >> 14;
-    t15a = (t11 - t15 + (1 << 13)) >> 14;
-
-    t4a  = t4 * 15137 + t5 *  6270;
-    t5a  = t4 *  6270 - t5 * 15137;
-    t6a  = t7 * 15137 - t6 *  6270;
-    t7a  = t7 *  6270 + t6 * 15137;
-    t12  = t12a * 15137 + t13a *  6270;
-    t13  = t12a *  6270 - t13a * 15137;
-    t14  = t15a * 15137 - t14a *  6270;
-    t15  = t15a *  6270 + t14a * 15137;
+    t8a  = (dctint)((1U << 13) + t8  + t12) >> 14;
+    t9a  = (dctint)((1U << 13) + t9  + t13) >> 14;
+    t10a = (dctint)((1U << 13) + t10 + t14) >> 14;
+    t11a = (dctint)((1U << 13) + t11 + t15) >> 14;
+    t12a = (dctint)((1U << 13) + t8  - t12) >> 14;
+    t13a = (dctint)((1U << 13) + t9  - t13) >> 14;
+    t14a = (dctint)((1U << 13) + t10 - t14) >> 14;
+    t15a = (dctint)((1U << 13) + t11 - t15) >> 14;
+
+    t4a  = t4 * 15137U + t5 *  6270U;
+    t5a  = t4 *  6270U - t5 * 15137U;
+    t6a  = t7 * 15137U - t6 *  6270U;
+    t7a  = t7 *  6270U + t6 * 15137U;
+    t12  = t12a * 15137U + t13a *  6270U;
+    t13  = t12a *  6270U - t13a * 15137U;
+    t14  = t15a * 15137U - t14a *  6270U;
+    t15  = t15a *  6270U + t14a * 15137U;
 
     out[ 0] =   t0 + t2;
     out[15] = -(t1 + t3);
     t2a     =   t0 - t2;
     t3a     =   t1 - t3;
-    out[ 3] = -((t4a + t6a + (1 << 13)) >> 14);
-    out[12] =   (t5a + t7a + (1 << 13)) >> 14;
-    t6      =   (t4a - t6a + (1 << 13)) >> 14;
-    t7      =   (t5a - t7a + (1 << 13)) >> 14;
+    out[ 3] = -((dctint)((1U << 13) + t4a + t6a) >> 14);
+    out[12] =   (dctint)((1U << 13) + t5a + t7a) >> 14;
+    t6      =   (dctint)((1U << 13) + t4a - t6a) >> 14;
+    t7      =   (dctint)((1U << 13) + t5a - t7a) >> 14;
     out[ 1] = -(t8a + t10a);
     out[14] =   t9a + t11a;
     t10     =   t8a - t10a;
     t11     =   t9a - t11a;
-    out[ 2] =   (t12 + t14 + (1 << 13)) >> 14;
-    out[13] = -((t13 + t15 + (1 << 13)) >> 14);
-    t14a    =   (t12 - t14 + (1 << 13)) >> 14;
-    t15a    =   (t13 - t15 + (1 << 13)) >> 14;
+    out[ 2] =   (dctint)((1U << 13) + t12 + t14) >> 14;
+    out[13] = -((dctint)((1U << 13) + t13 + t15) >> 14);
+    t14a    =   (dctint)((1U << 13) + t12 - t14) >> 14;
+    t15a    =   (dctint)((1U << 13) + t13 - t15) >> 14;
 
-    out[ 7] = ((t2a  + t3a)  * -11585 + (1 << 13)) >> 14;
-    out[ 8] = ((t2a  - t3a)  *  11585 + (1 << 13)) >> 14;
-    out[ 4] = ((t7   + t6)   *  11585 + (1 << 13)) >> 14;
-    out[11] = ((t7   - t6)   *  11585 + (1 << 13)) >> 14;
-    out[ 6] = ((t11  + t10)  *  11585 + (1 << 13)) >> 14;
-    out[ 9] = ((t11  - t10)  *  11585 + (1 << 13)) >> 14;
-    out[ 5] = ((t14a + t15a) * -11585 + (1 << 13)) >> 14;
-    out[10] = ((t14a - t15a) *  11585 + (1 << 13)) >> 14;
+    out[ 7] = (dctint)(-(t2a  + t3a)  * 11585U  + (1 << 13)) >> 14;
+    out[ 8] = (dctint)( (t2a  - t3a)  * 11585U  + (1 << 13)) >> 14;
+    out[ 4] = (dctint)( (t7   + t6)   * 11585U  + (1 << 13)) >> 14;
+    out[11] = (dctint)( (t7   - t6)   * 11585U  + (1 << 13)) >> 14;
+    out[ 6] = (dctint)( (t11  + t10)  * 11585U  + (1 << 13)) >> 14;
+    out[ 9] = (dctint)( (t11  - t10)  * 11585U  + (1 << 13)) >> 14;
+    out[ 5] = (dctint)(-(t14a + t15a) * 11585U  + (1 << 13)) >> 14;
+    out[10] = (dctint)( (t14a - t15a) * 11585U  + (1 << 13)) >> 14;
 }
 
 itxfm_wrap(16, 6)
@@ -1479,38 +1479,38 @@ itxfm_wrap(16, 6)
 static av_always_inline void idct32_1d(const dctcoef *in, ptrdiff_t stride,
                                        dctcoef *out, int pass)
 {
-    dctint t0a  = ((IN(0) + IN(16)) * 11585 + (1 << 13)) >> 14;
-    dctint t1a  = ((IN(0) - IN(16)) * 11585 + (1 << 13)) >> 14;
-    dctint t2a  = (IN( 8) *  6270 - IN(24) * 15137 + (1 << 13)) >> 14;
-    dctint t3a  = (IN( 8) * 15137 + IN(24) *  6270 + (1 << 13)) >> 14;
-    dctint t4a  = (IN( 4) *  3196 - IN(28) * 16069 + (1 << 13)) >> 14;
-    dctint t7a  = (IN( 4) * 16069 + IN(28) *  3196 + (1 << 13)) >> 14;
-    dctint t5a  = (IN(20) * 13623 - IN(12) *  9102 + (1 << 13)) >> 14;
-    dctint t6a  = (IN(20) *  9102 + IN(12) * 13623 + (1 << 13)) >> 14;
-    dctint t8a  = (IN( 2) *  1606 - IN(30) * 16305 + (1 << 13)) >> 14;
-    dctint t15a = (IN( 2) * 16305 + IN(30) *  1606 + (1 << 13)) >> 14;
-    dctint t9a  = (IN(18) * 12665 - IN(14) * 10394 + (1 << 13)) >> 14;
-    dctint t14a = (IN(18) * 10394 + IN(14) * 12665 + (1 << 13)) >> 14;
-    dctint t10a = (IN(10) *  7723 - IN(22) * 14449 + (1 << 13)) >> 14;
-    dctint t13a = (IN(10) * 14449 + IN(22) *  7723 + (1 << 13)) >> 14;
-    dctint t11a = (IN(26) * 15679 - IN( 6) *  4756 + (1 << 13)) >> 14;
-    dctint t12a = (IN(26) *  4756 + IN( 6) * 15679 + (1 << 13)) >> 14;
-    dctint t16a = (IN( 1) *   804 - IN(31) * 16364 + (1 << 13)) >> 14;
-    dctint t31a = (IN( 1) * 16364 + IN(31) *   804 + (1 << 13)) >> 14;
-    dctint t17a = (IN(17) * 12140 - IN(15) * 11003 + (1 << 13)) >> 14;
-    dctint t30a = (IN(17) * 11003 + IN(15) * 12140 + (1 << 13)) >> 14;
-    dctint t18a = (IN( 9) *  7005 - IN(23) * 14811 + (1 << 13)) >> 14;
-    dctint t29a = (IN( 9) * 14811 + IN(23) *  7005 + (1 << 13)) >> 14;
-    dctint t19a = (IN(25) * 15426 - IN( 7) *  5520 + (1 << 13)) >> 14;
-    dctint t28a = (IN(25) *  5520 + IN( 7) * 15426 + (1 << 13)) >> 14;
-    dctint t20a = (IN( 5) *  3981 - IN(27) * 15893 + (1 << 13)) >> 14;
-    dctint t27a = (IN( 5) * 15893 + IN(27) *  3981 + (1 << 13)) >> 14;
-    dctint t21a = (IN(21) * 14053 - IN(11) *  8423 + (1 << 13)) >> 14;
-    dctint t26a = (IN(21) *  8423 + IN(11) * 14053 + (1 << 13)) >> 14;
-    dctint t22a = (IN(13) *  9760 - IN(19) * 13160 + (1 << 13)) >> 14;
-    dctint t25a = (IN(13) * 13160 + IN(19) *  9760 + (1 << 13)) >> 14;
-    dctint t23a = (IN(29) * 16207 - IN( 3) *  2404 + (1 << 13)) >> 14;
-    dctint t24a = (IN(29) *  2404 + IN( 3) * 16207 + (1 << 13)) >> 14;
+    dctint t0a  = (dctint)((IN(0) + IN(16)) * 11585U         + (1 << 13)) >> 14;
+    dctint t1a  = (dctint)((IN(0) - IN(16)) * 11585U         + (1 << 13)) >> 14;
+    dctint t2a  = (dctint)(IN( 8) *  6270U - IN(24) * 15137U + (1 << 13)) >> 14;
+    dctint t3a  = (dctint)(IN( 8) * 15137U + IN(24) *  6270U + (1 << 13)) >> 14;
+    dctint t4a  = (dctint)(IN( 4) *  3196U - IN(28) * 16069U + (1 << 13)) >> 14;
+    dctint t7a  = (dctint)(IN( 4) * 16069U + IN(28) *  3196U + (1 << 13)) >> 14;
+    dctint t5a  = (dctint)(IN(20) * 13623U - IN(12) *  9102U + (1 << 13)) >> 14;
+    dctint t6a  = (dctint)(IN(20) *  9102U + IN(12) * 13623U + (1 << 13)) >> 14;
+    dctint t8a  = (dctint)(IN( 2) *  1606U - IN(30) * 16305U + (1 << 13)) >> 14;
+    dctint t15a = (dctint)(IN( 2) * 16305U + IN(30) *  1606U + (1 << 13)) >> 14;
+    dctint t9a  = (dctint)(IN(18) * 12665U - IN(14) * 10394U + (1 << 13)) >> 14;
+    dctint t14a = (dctint)(IN(18) * 10394U + IN(14) * 12665U + (1 << 13)) >> 14;
+    dctint t10a = (dctint)(IN(10) *  7723U - IN(22) * 14449U + (1 << 13)) >> 14;
+    dctint t13a = (dctint)(IN(10) * 14449U + IN(22) *  7723U + (1 << 13)) >> 14;
+    dctint t11a = (dctint)(IN(26) * 15679U - IN( 6) *  4756U + (1 << 13)) >> 14;
+    dctint t12a = (dctint)(IN(26) *  4756U + IN( 6) * 15679U + (1 << 13)) >> 14;
+    dctint t16a = (dctint)(IN( 1) *   804U - IN(31) * 16364U + (1 << 13)) >> 14;
+    dctint t31a = (dctint)(IN( 1) * 16364U + IN(31) *   804U + (1 << 13)) >> 14;
+    dctint t17a = (dctint)(IN(17) * 12140U - IN(15) * 11003U + (1 << 13)) >> 14;
+    dctint t30a = (dctint)(IN(17) * 11003U + IN(15) * 12140U + (1 << 13)) >> 14;
+    dctint t18a = (dctint)(IN( 9) *  7005U - IN(23) * 14811U + (1 << 13)) >> 14;
+    dctint t29a = (dctint)(IN( 9) * 14811U + IN(23) *  7005U + (1 << 13)) >> 14;
+    dctint t19a = (dctint)(IN(25) * 15426U - IN( 7) *  5520U + (1 << 13)) >> 14;
+    dctint t28a = (dctint)(IN(25) *  5520U + IN( 7) * 15426U + (1 << 13)) >> 14;
+    dctint t20a = (dctint)(IN( 5) *  3981U - IN(27) * 15893U + (1 << 13)) >> 14;
+    dctint t27a = (dctint)(IN( 5) * 15893U + IN(27) *  3981U + (1 << 13)) >> 14;
+    dctint t21a = (dctint)(IN(21) * 14053U - IN(11) *  8423U + (1 << 13)) >> 14;
+    dctint t26a = (dctint)(IN(21) *  8423U + IN(11) * 14053U + (1 << 13)) >> 14;
+    dctint t22a = (dctint)(IN(13) *  9760U - IN(19) * 13160U + (1 << 13)) >> 14;
+    dctint t25a = (dctint)(IN(13) * 13160U + IN(19) *  9760U + (1 << 13)) >> 14;
+    dctint t23a = (dctint)(IN(29) * 16207U - IN( 3) *  2404U + (1 << 13)) >> 14;
+    dctint t24a = (dctint)(IN(29) *  2404U + IN( 3) * 16207U + (1 << 13)) >> 14;
 
     dctint t0  = t0a  + t3a;
     dctint t1  = t1a  + t2a;
@@ -1545,20 +1545,20 @@ static av_always_inline void idct32_1d(const dctcoef *in, ptrdiff_t stride,
     dctint t30 = t31a - t30a;
     dctint t31 = t31a + t30a;
 
-    t5a = ((t6 - t5) * 11585 + (1 << 13)) >> 14;
-    t6a = ((t6 + t5) * 11585 + (1 << 13)) >> 14;
-    t9a  = (  t14 *  6270 - t9  * 15137  + (1 << 13)) >> 14;
-    t14a = (  t14 * 15137 + t9  *  6270  + (1 << 13)) >> 14;
-    t10a = (-(t13 * 15137 + t10 *  6270) + (1 << 13)) >> 14;
-    t13a = (  t13 *  6270 - t10 * 15137  + (1 << 13)) >> 14;
-    t17a = (  t30 *  3196 - t17 * 16069  + (1 << 13)) >> 14;
-    t30a = (  t30 * 16069 + t17 *  3196  + (1 << 13)) >> 14;
-    t18a = (-(t29 * 16069 + t18 *  3196) + (1 << 13)) >> 14;
-    t29a = (  t29 *  3196 - t18 * 16069  + (1 << 13)) >> 14;
-    t21a = (  t26 * 13623 - t21 *  9102  + (1 << 13)) >> 14;
-    t26a = (  t26 *  9102 + t21 * 13623  + (1 << 13)) >> 14;
-    t22a = (-(t25 *  9102 + t22 * 13623) + (1 << 13)) >> 14;
-    t25a = (  t25 * 13623 - t22 *  9102  + (1 << 13)) >> 14;
+    t5a  = (dctint)((t6 - t5) * 11585U             + (1 << 13)) >> 14;
+    t6a  = (dctint)((t6 + t5) * 11585U             + (1 << 13)) >> 14;
+    t9a  = (dctint)(  t14 *  6270U - t9  * 15137U  + (1 << 13)) >> 14;
+    t14a = (dctint)(  t14 * 15137U + t9  *  6270U  + (1 << 13)) >> 14;
+    t10a = (dctint)(-(t13 * 15137U + t10 *  6270U) + (1 << 13)) >> 14;
+    t13a = (dctint)(  t13 *  6270U - t10 * 15137U  + (1 << 13)) >> 14;
+    t17a = (dctint)(  t30 *  3196U - t17 * 16069U  + (1 << 13)) >> 14;
+    t30a = (dctint)(  t30 * 16069U + t17 *  3196U  + (1 << 13)) >> 14;
+    t18a = (dctint)(-(t29 * 16069U + t18 *  3196U) + (1 << 13)) >> 14;
+    t29a = (dctint)(  t29 *  3196U - t18 * 16069U  + (1 << 13)) >> 14;
+    t21a = (dctint)(  t26 * 13623U - t21 *  9102U  + (1 << 13)) >> 14;
+    t26a = (dctint)(  t26 *  9102U + t21 * 13623U  + (1 << 13)) >> 14;
+    t22a = (dctint)(-(t25 *  9102U + t22 * 13623U) + (1 << 13)) >> 14;
+    t25a = (dctint)(  t25 * 13623U - t22 *  9102U  + (1 << 13)) >> 14;
 
     t0a  = t0   + t7;
     t1a  = t1   + t6a;
@@ -1593,18 +1593,18 @@ static av_always_inline void idct32_1d(const dctcoef *in, ptrdiff_t stride,
     t30  = t30a + t29a;
     t31a = t31  + t28;
 
-    t10a = ((t13  - t10)  * 11585 + (1 << 13)) >> 14;
-    t13a = ((t13  + t10)  * 11585 + (1 << 13)) >> 14;
-    t11  = ((t12a - t11a) * 11585 + (1 << 13)) >> 14;
-    t12  = ((t12a + t11a) * 11585 + (1 << 13)) >> 14;
-    t18a = (  t29  *  6270 - t18  * 15137  + (1 << 13)) >> 14;
-    t29a = (  t29  * 15137 + t18  *  6270  + (1 << 13)) >> 14;
-    t19  = (  t28a *  6270 - t19a * 15137  + (1 << 13)) >> 14;
-    t28  = (  t28a * 15137 + t19a *  6270  + (1 << 13)) >> 14;
-    t20  = (-(t27a * 15137 + t20a *  6270) + (1 << 13)) >> 14;
-    t27  = (  t27a *  6270 - t20a * 15137  + (1 << 13)) >> 14;
-    t21a = (-(t26  * 15137 + t21  *  6270) + (1 << 13)) >> 14;
-    t26a = (  t26  *  6270 - t21  * 15137  + (1 << 13)) >> 14;
+    t10a = (dctint)((t13  - t10)  * 11585U           + (1 << 13)) >> 14;
+    t13a = (dctint)((t13  + t10)  * 11585U           + (1 << 13)) >> 14;
+    t11  = (dctint)((t12a - t11a) * 11585U           + (1 << 13)) >> 14;
+    t12  = (dctint)((t12a + t11a) * 11585U           + (1 << 13)) >> 14;
+    t18a = (dctint)(  t29  *  6270U - t18  * 15137U  + (1 << 13)) >> 14;
+    t29a = (dctint)(  t29  * 15137U + t18  *  6270U  + (1 << 13)) >> 14;
+    t19  = (dctint)(  t28a *  6270U - t19a * 15137U  + (1 << 13)) >> 14;
+    t28  = (dctint)(  t28a * 15137U + t19a *  6270U  + (1 << 13)) >> 14;
+    t20  = (dctint)(-(t27a * 15137U + t20a *  6270U) + (1 << 13)) >> 14;
+    t27  = (dctint)(  t27a *  6270U - t20a * 15137U  + (1 << 13)) >> 14;
+    t21a = (dctint)(-(t26  * 15137U + t21  *  6270U) + (1 << 13)) >> 14;
+    t26a = (dctint)(  t26  *  6270U - t21  * 15137U  + (1 << 13)) >> 14;
 
     t0   = t0a + t15a;
     t1   = t1a + t14;
@@ -1639,14 +1639,14 @@ static av_always_inline void idct32_1d(const dctcoef *in, ptrdiff_t stride,
     t30a = t30  + t25;
     t31  = t31a + t24a;
 
-    t20  = ((t27a - t20a) * 11585 + (1 << 13)) >> 14;
-    t27  = ((t27a + t20a) * 11585 + (1 << 13)) >> 14;
-    t21a = ((t26  - t21 ) * 11585 + (1 << 13)) >> 14;
-    t26a = ((t26  + t21 ) * 11585 + (1 << 13)) >> 14;
-    t22  = ((t25a - t22a) * 11585 + (1 << 13)) >> 14;
-    t25  = ((t25a + t22a) * 11585 + (1 << 13)) >> 14;
-    t23a = ((t24  - t23 ) * 11585 + (1 << 13)) >> 14;
-    t24a = ((t24  + t23 ) * 11585 + (1 << 13)) >> 14;
+    t20  = (dctint)((t27a - t20a) * 11585U + (1 << 13)) >> 14;
+    t27  = (dctint)((t27a + t20a) * 11585U + (1 << 13)) >> 14;
+    t21a = (dctint)((t26  - t21 ) * 11585U + (1 << 13)) >> 14;
+    t26a = (dctint)((t26  + t21 ) * 11585U + (1 << 13)) >> 14;
+    t22  = (dctint)((t25a - t22a) * 11585U + (1 << 13)) >> 14;
+    t25  = (dctint)((t25a + t22a) * 11585U + (1 << 13)) >> 14;
+    t23a = (dctint)((t24  - t23 ) * 11585U + (1 << 13)) >> 14;
+    t24a = (dctint)((t24  + t23 ) * 11585U + (1 << 13)) >> 14;
 
     out[ 0] = t0   + t31;
     out[ 1] = t1   + t30a;